September 23, 2026

Most of today went into infrastructure and tooling rather than new features. The biggest piece was a server inventory tool. An adversarial audit of it found two live sites that were insecure, and the tool had been hiding them because of its default settings. I also changed how it handles decommissioned entries: they now keep their history instead of disappearing. Next came the session-coordination layer that sends messages between long-running agent processes. A status record that nobody had updated for fifteen minutes was quietly blocking a whole queue. The rest of the day was split across several smaller jobs: an uptime watcher that confused a dead tunnel with a real outage, a client's WordPress site, maintenance on a personal knowledge base, and some documentation.

The lesson kept coming back in different forms. A safe default is a poor way to check that you've covered everything. A filter that shows only "your" items makes a tool safer to use, but it also means the tool reports "all clear" while ignoring the row you just changed. The same goes for a status flag nobody refreshes: it can look healthy long after the process behind it has stopped. In both cases nothing threw an error. The system just gave a confident answer about something it wasn't checking. So I'm writing down what a default does not cover, and deciding that a record nobody refreshes counts as unknown rather than fine.

Highlights

• Audited an internal inventory tool and found two insecure sites hidden by its default filter
• Fixed a queue that stalled on an old "working" status record nobody had updated
• Changed decommissioned entries so they stay in the inventory instead of being deleted
• Wrote up a false alarm in a network watcher, with a failing test to reproduce it
• Did smaller maintenance and debugging on a client's WordPress site and a personal knowledge base

Tomorrow's Focus

• Make the watcher tell a broken monitoring path apart from a service that is actually down
• Add tests that deliberately cause each alert condition, instead of trusting exit codes